Skip to content

Commit f96e85b

Browse files
mariusz-sikora-at-amdmariusz-sikora-at-amd
authored
[AMDGPU][GFX12] Add tests for unsupported builtins (#78729)
__builtin_amdgcn_mfma* and __builtin_amdgcn_smfmac*
1 parent e3c9327 commit f96e85b

File tree

1 file changed

+85
-1
lines changed

1 file changed

+85
-1
lines changed

‎clang/test/CodeGenOpenCL/builtins-amdgcn-gfx12-err.cl

Lines changed: 85 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -4,10 +4,94 @@
44

55
typedef unsigned int uint;
66

7-
kernel void test_builtins_amdgcn_gws_insts(uint a, uint b) {
7+
#pragma OPENCL EXTENSION cl_khr_fp64:enable
8+
9+
typedef float v2f __attribute__((ext_vector_type(2)));
10+
typedef float v4f __attribute__((ext_vector_type(4)));
11+
typedef float v16f __attribute__((ext_vector_type(16)));
12+
typedef float v32f __attribute__((ext_vector_type(32)));
13+
typedef half v4h __attribute__((ext_vector_type(4)));
14+
typedef half v8h __attribute__((ext_vector_type(8)));
15+
typedef half v16h __attribute__((ext_vector_type(16)));
16+
typedef half v32h __attribute__((ext_vector_type(32)));
17+
typedef int v2i __attribute__((ext_vector_type(2)));
18+
typedef int v4i __attribute__((ext_vector_type(4)));
19+
typedef int v16i __attribute__((ext_vector_type(16)));
20+
typedef int v32i __attribute__((ext_vector_type(32)));
21+
typedef short v2s __attribute__((ext_vector_type(2)));
22+
typedef short v4s __attribute__((ext_vector_type(4)));
23+
typedef short v8s __attribute__((ext_vector_type(8)));
24+
typedef short v16s __attribute__((ext_vector_type(16)));
25+
typedef short v32s __attribute__((ext_vector_type(32)));
26+
typedef double v4d __attribute__((ext_vector_type(4)));
27+
28+
void builtin_test_unsupported(double a_double, float a_float,
29+
int a_int, long a_long,
30+
v4d a_v4d,
31+
v2s a_v2s, v4s a_v4s, v8s a_v8s,
32+
v2i a_v2i, v4i a_v4i, v16i a_v16i, v32i a_v32i,
33+
v2f a_v2f, v4f a_v4f, v16f a_v16f, v32f a_v32f,
34+
v4h a_v4h, v8h a_v8h,
35+
36+
uint a, uint b) {
37+
838
__builtin_amdgcn_ds_gws_init(a, b); // expected-error {{'__builtin_amdgcn_ds_gws_init' needs target feature gws}}
939
__builtin_amdgcn_ds_gws_barrier(a, b); // expected-error {{'__builtin_amdgcn_ds_gws_barrier' needs target feature gws}}
1040
__builtin_amdgcn_ds_gws_sema_v(a); // expected-error {{'__builtin_amdgcn_ds_gws_sema_v' needs target feature gws}}
1141
__builtin_amdgcn_ds_gws_sema_br(a, b); // expected-error {{'__builtin_amdgcn_ds_gws_sema_br' needs target feature gws}}
1242
__builtin_amdgcn_ds_gws_sema_p(a); // expected-error {{'__builtin_amdgcn_ds_gws_sema_p' needs target feature gws}}
43+
44+
a_v32f = __builtin_amdgcn_mfma_f32_32x32x1f32(a_float, a_float, a_v32f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x1f32' needs target feature mai-insts}}
45+
a_v16f = __builtin_amdgcn_mfma_f32_16x16x1f32(a_float, a_float, a_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x1f32' needs target feature mai-insts}}
46+
a_v4f = __builtin_amdgcn_mfma_f32_4x4x1f32(a_float, a_float, a_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_4x4x1f32' needs target feature mai-insts}}
47+
a_v16f = __builtin_amdgcn_mfma_f32_32x32x2f32(a_float, a_float, a_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x2f32' needs target feature mai-insts}}
48+
a_v4f = __builtin_amdgcn_mfma_f32_16x16x4f32(a_float, a_float, a_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x4f32' needs target feature mai-insts}}
49+
a_v32f = __builtin_amdgcn_mfma_f32_32x32x4f16(a_v4h, a_v4h, a_v32f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x4f16' needs target feature mai-insts}}
50+
a_v16f = __builtin_amdgcn_mfma_f32_16x16x4f16(a_v4h, a_v4h, a_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x4f16' needs target feature mai-insts}}
51+
a_v4f = __builtin_amdgcn_mfma_f32_4x4x4f16(a_v4h, a_v4h, a_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_4x4x4f16' needs target feature mai-insts}}
52+
a_v16f = __builtin_amdgcn_mfma_f32_32x32x8f16(a_v4h, a_v4h, a_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x8f16' needs target feature mai-insts}}
53+
a_v4f = __builtin_amdgcn_mfma_f32_16x16x16f16(a_v4h, a_v4h, a_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x16f16' needs target feature mai-insts}}
54+
a_v32i = __builtin_amdgcn_mfma_i32_32x32x4i8(a_int, a_int, a_v32i, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_i32_32x32x4i8' needs target feature mai-insts}}
55+
a_v16i = __builtin_amdgcn_mfma_i32_16x16x4i8(a_int, a_int, a_v16i, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_i32_16x16x4i8' needs target feature mai-insts}}
56+
a_v4i = __builtin_amdgcn_mfma_i32_4x4x4i8(a_int, a_int, a_v4i, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_i32_4x4x4i8' needs target feature mai-insts}}
57+
a_v16i = __builtin_amdgcn_mfma_i32_32x32x8i8(a_int, a_int, a_v16i, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_i32_32x32x8i8' needs target feature mai-insts}}
58+
a_v4i = __builtin_amdgcn_mfma_i32_16x16x16i8(a_int, a_int, a_v4i, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_i32_16x16x16i8' needs target feature mai-insts}}
59+
a_v32f = __builtin_amdgcn_mfma_f32_32x32x2bf16(a_v2s, a_v2s, a_v32f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x2bf16' needs target feature mai-insts}}
60+
a_v16f = __builtin_amdgcn_mfma_f32_16x16x2bf16(a_v2s, a_v2s, a_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x2bf16' needs target feature mai-insts}}
61+
a_v4f = __builtin_amdgcn_mfma_f32_4x4x2bf16(a_v2s, a_v2s, a_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_4x4x2bf16' needs target feature mai-insts}}
62+
a_v16f = __builtin_amdgcn_mfma_f32_32x32x4bf16(a_v2s, a_v2s, a_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x4bf16' needs target feature mai-insts}}
63+
a_v4f = __builtin_amdgcn_mfma_f32_16x16x8bf16(a_v2s, a_v2s, a_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x8bf16' needs target feature mai-insts}}
64+
a_v32f = __builtin_amdgcn_mfma_f32_32x32x4bf16_1k(a_v4s, a_v4s, a_v32f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x4bf16_1k' needs target feature mai-insts}}
65+
a_v16f = __builtin_amdgcn_mfma_f32_16x16x4bf16_1k(a_v4s, a_v4s, a_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x4bf16_1k' needs target feature mai-insts}}
66+
a_v4f = __builtin_amdgcn_mfma_f32_4x4x4bf16_1k(a_v4s, a_v4s, a_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_4x4x4bf16_1k' needs target feature mai-insts}}
67+
a_v16f = __builtin_amdgcn_mfma_f32_32x32x8bf16_1k(a_v4s, a_v4s, a_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x8bf16_1k' needs target feature mai-insts}}
68+
a_v4f = __builtin_amdgcn_mfma_f32_16x16x16bf16_1k(a_v4s, a_v4s, a_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x16bf16_1k' needs target feature mai-insts}}
69+
a_v4d = __builtin_amdgcn_mfma_f64_16x16x4f64(a_double, a_double, a_v4d, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f64_16x16x4f64' needs target feature mai-insts}}
70+
a_double = __builtin_amdgcn_mfma_f64_4x4x4f64(a_double, a_double, a_double, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f64_4x4x4f64' needs target feature mai-insts}}
71+
a_v4i = __builtin_amdgcn_mfma_i32_16x16x32_i8(a_long, a_long, a_v4i, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_i32_16x16x32_i8' needs target feature mai-insts}}
72+
a_v16i = __builtin_amdgcn_mfma_i32_32x32x16_i8(a_long, a_long, a_v16i, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_i32_32x32x16_i8' needs target feature mai-insts}}
73+
a_v4f = __builtin_amdgcn_mfma_f32_16x16x8_xf32(a_v2f, a_v2f, a_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x8_xf32' needs target feature mai-insts}}
74+
a_v16f = __builtin_amdgcn_mfma_f32_32x32x4_xf32(a_v2f, a_v2f, a_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x4_xf32' needs target feature mai-insts}}
75+
a_v4f = __builtin_amdgcn_mfma_f32_16x16x32_bf8_bf8(a_long, a_long, a_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x32_bf8_bf8' needs target feature fp8-insts}}
76+
a_v4f = __builtin_amdgcn_mfma_f32_16x16x32_bf8_fp8(a_long, a_long, a_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x32_bf8_fp8' needs target feature fp8-insts}}
77+
a_v4f = __builtin_amdgcn_mfma_f32_16x16x32_fp8_bf8(a_long, a_long, a_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x32_fp8_bf8' needs target feature fp8-insts}}
78+
a_v4f = __builtin_amdgcn_mfma_f32_16x16x32_fp8_fp8(a_long, a_long, a_v4f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_16x16x32_fp8_fp8' needs target feature fp8-insts}}
79+
a_v16f = __builtin_amdgcn_mfma_f32_32x32x16_bf8_bf8(a_long, a_long, a_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x16_bf8_bf8' needs target feature fp8-insts}}
80+
a_v16f = __builtin_amdgcn_mfma_f32_32x32x16_bf8_fp8(a_long, a_long, a_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x16_bf8_fp8' needs target feature fp8-insts}}
81+
a_v16f = __builtin_amdgcn_mfma_f32_32x32x16_fp8_bf8(a_long, a_long, a_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x16_fp8_bf8' needs target feature fp8-insts}}
82+
a_v16f = __builtin_amdgcn_mfma_f32_32x32x16_fp8_fp8(a_long, a_long, a_v16f, 0, 0, 0); // expected-error {{'__builtin_amdgcn_mfma_f32_32x32x16_fp8_fp8' needs target feature fp8-insts}}
83+
a_v4f = __builtin_amdgcn_smfmac_f32_16x16x32_f16(a_v4h, a_v8h, a_v4f, a_int, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_16x16x32_f16' needs target feature mai-insts}}
84+
a_v16f = __builtin_amdgcn_smfmac_f32_32x32x16_f16(a_v4h, a_v8h, a_v16f, a_int, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_32x32x16_f16' needs target feature mai-insts}}
85+
a_v4f = __builtin_amdgcn_smfmac_f32_16x16x32_bf16(a_v4s, a_v8s, a_v4f, a_int, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_16x16x32_bf16' needs target feature mai-insts}}
86+
a_v16f = __builtin_amdgcn_smfmac_f32_32x32x16_bf16(a_v4s, a_v8s, a_v16f, a_int, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_32x32x16_bf16' needs target feature mai-insts}}
87+
a_v4i = __builtin_amdgcn_smfmac_i32_16x16x64_i8(a_v2i, a_v4i, a_v4i, a_int, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_i32_16x16x64_i8' needs target feature mai-insts}}
88+
a_v16i = __builtin_amdgcn_smfmac_i32_32x32x32_i8(a_v2i, a_v4i, a_v16i, a_int, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_i32_32x32x32_i8' needs target feature mai-insts}}
89+
a_v4f = __builtin_amdgcn_smfmac_f32_16x16x64_bf8_bf8(a_v2i, a_v4i, a_v4f, a_int, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_16x16x64_bf8_bf8' needs target feature fp8-insts}}
90+
a_v4f = __builtin_amdgcn_smfmac_f32_16x16x64_bf8_fp8(a_v2i, a_v4i, a_v4f, a_int, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_16x16x64_bf8_fp8' needs target feature fp8-insts}}
91+
a_v4f = __builtin_amdgcn_smfmac_f32_16x16x64_fp8_bf8(a_v2i, a_v4i, a_v4f, a_int, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_16x16x64_fp8_bf8' needs target feature fp8-insts}}
92+
a_v4f = __builtin_amdgcn_smfmac_f32_16x16x64_fp8_fp8(a_v2i, a_v4i, a_v4f, a_int, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_16x16x64_fp8_fp8' needs target feature fp8-insts}}
93+
a_v16f = __builtin_amdgcn_smfmac_f32_32x32x32_bf8_bf8(a_v2i, a_v4i, a_v16f, a_int, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_32x32x32_bf8_bf8' needs target feature fp8-insts}}
94+
a_v16f = __builtin_amdgcn_smfmac_f32_32x32x32_bf8_fp8(a_v2i, a_v4i, a_v16f, a_int, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_32x32x32_bf8_fp8' needs target feature fp8-insts}}
95+
a_v16f = __builtin_amdgcn_smfmac_f32_32x32x32_fp8_bf8(a_v2i, a_v4i, a_v16f, a_int, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_32x32x32_fp8_bf8' needs target feature fp8-insts}}
96+
a_v16f = __builtin_amdgcn_smfmac_f32_32x32x32_fp8_fp8(a_v2i, a_v4i, a_v16f, a_int, 0, 0); // expected-error {{'__builtin_amdgcn_smfmac_f32_32x32x32_fp8_fp8' needs target feature fp8-insts}}
1397
}

0 commit comments

Comments
 (0)